Identification of Extreme Capital Flows in Emerging Markets

Transcription

1 Identification of Extreme Capital Flows in Emerging Markets Amrita Dhar February 2016 Abstract Capital flows to emerging market economies can be characterized by periods of large inflows alternating with periods of large outflows. Exceptionally high levels of flows have often been associated with financial crises, and identifying such episodes is crucial for understanding the onset of crises. The existing literature, however, relies on ad hoc threshold criteria for identifying such extreme episodes. This paper identifies extreme episodes from the data using a formal statistical classification. In particular, I employ a three state Markov-switching model to characterize extreme episodes of quarterly net capital flows for each country in a sample of 36 emerging market economies from 1980 to The model identifies 7.8 percent of the total sample as periods of extreme inflows ( surges ) and 3 percent of the total sample as extreme outflows ( flights ). Compared to the literature, the model identifies fewer episodes as extreme, and the number of episodes varies substantially across countries. JEL classification: F21,F32, F39 Keywords: Capital flows, Surges, Flights, Emerging market economies Department of Economics, University of Houston, Houston, TX 77204, adhar2@uh.edu I am grateful to my advisors David Papell and Bent Sørensen for their valuable comments and continued support throughout the project. I thank Dietrich Vollrath, German Cubas, Liliana Varela, and Christopher Biolsi for their insightful comments and suggestions. I also thank the participants of the 80th International Atlantic Economic Conference as well as the Brown Bag, and Graduate Student Workshops attendees at the University of Houston for their helpful comments. All errors are my own. 1

2 1 Introduction Extreme capital flows have often posed challenges for policy makers in emerging markets (see Calvo, 1998; Edwards, 2000; Hutchison and Noy, 2006). A sudden large inflow of capital puts pressure on the domestic currency causing it to appreciate which in turn has an adverse effect on the country s net exports performance. Analogously, a sudden large outflow of capital causes the domestic currency to depreciate, often steeply so, and cause inflationary pressure. 1 After the Great Recession, interest rates hit the zero lower bound for many developed countries, causing capital to flow to countries with higher returns, in particular emerging markets. This led to a resurgence of interests among academics as well as policy makers in understanding flows of capital and their consequences in these economies (see Eichengreen and Gupta, 2014; Powell, 2013; Ahmed and Zlate, 2013). In December 2012, the International Monetary Fund (IMF) adopted a new institutional view on capital account liberalization and management of capital flows, and the IMF now acknowledges that regulation of capital flows is desirable under certain circumstances (see IMF Executive Summary Report, 2012). There is an extensive literature studying ebbs and flows of capital to emerging markets going back to Calvo (1998), who documents episodes of sudden stops, using current account data as a proxy for capital flows, as abrupt declines in capital flows. Reinhart and Reinhart (2009) talk about capital flow bonanzas, i.e., when economies experience a huge influx of capital. Part of their analysis involves dating such episodes for a large sample of countries. 2 More recently, Forbes and Warnock (2012) and Ghosh et al. (2014) study actual capital flows and analyze periods of extreme flows. The focus among economists so far has mostly been on analyzing either the causes or the consequences of extreme episodes. Clearly, identification of extreme episodes is crucial for doing such analysis. The existing literature sets ad hoc threshold criteria for identifying these episodes. 3 My objective is to take a step back and focus on rigorous identification of the capital flow episodes using a formal statistical model. This paper is, to my knowledge, the first study to conduct a formal statistical classification of extreme capital flow episodes. 1 Blanchard et al. (2015) have showed these flows can have positive effects as well depending on the nature of flows. 2 Several other studies look into surges in capital flows. See, for example, Caballero (2012), Powell and Tavella (2015), and Cardarelli, Elekdag, and Kose (2010). 3 See Reinhart and Reinhart, 2009; Cardarelli, Elekdag, and Kose, 2010; Ghosh et al., 2014;?. 2

3 I apply a regime switching model, as suggested by Hamilton (1989) for classifying business cycles in U.S. GDP, to characterize capital flow episodes for a sample of emerging markets. This nonlinear model captures switches between different regimes and characterizes average flows in each regime. I use a three state Markov-switching model allowing the mean of the magnitude of the flows to switch between states of extreme, high, and low flows for each country. The switches between different states in this model are treated as inherent to the data-generating process, which allows the extreme episodes to be determined endogenously from the data. Going forward, I will refer to the states and regimes interchangeably. Using quarterly net capital flows, as a percentage of GDP, for a sample of 36 emerging markets from 1980 to 2014, the model identifies a total of 289 quarters of extreme net inflow periods or surges (around 7.8% of the total observations in the sample and 110 quarters of extreme net outflow episodes, or flights, (about 3% of the total observations in the sample). The number of quarters in extreme states varies considerably across countries. I estimate the model for each country at a time. I find on an average, the mean of net capital inflows during the surges is four times the mean in low inflow episodes and around double the mean of flows during the high inflow regimes. The mean value of net outflows during flights is six times the value of the net flows of low outflows and almost thrice the value under high outflow regimes. In comparison to the existing threshold criteria used in the literature, which generally labels the top 20 to 35 percent of the net capital flow distributions as extreme, the Markov-switching model identifies fewer episodes. Even restricting the threshold criteria to match the number of surges that the Markov-switching model identifies, the surge episodes identified under the two methods are not generally the same. This is because the Markov switching model identifies regimes rather than observations, where the identification of an observation as extreme depends on the pattern of observations nearby in time. Hence, the model may pick half the sample, or just 1 percent of the sample as extreme episodes depending on the size of the flows relative to the variation in other periods. The paper is organized as follows. A brief review of the literature is provided in Section 2. In Section 3, I briefly describe the data and explain the methodology. Section 4 analyzes the results. In Section 5, I compare the episodes identified using my methodology with the episodes identified by existing methodologies in the literature, and finally Section 6 concludes. 3

4 2 Literature Review on Capital Flows The literature on extreme capital flow goes back to Calvo (1998) who provides a theoretical analysis of the consequences of sudden stops of private capital flows to emerging market nations. He defines sudden stops as sharp unexpected contractions in net capital flows using country s current account deficits as a proxy. In a later study, Calvo, Izquierdo, and Mejia (2004) conduct an empirical analysis of the sudden stops. The definition of sudden stops they use for the empirical analysis is based on a threshold approach. In particular, according to their definition, a sudden stop phase begins when the annual change in the current account deficit is one standard deviation below its sample mean and ends when it goes above one standard deviation above the mean. It has to go down two standard deviation below the mean during the phase in order for the entire phase to be considered as a sudden stop. They further impose an additional criterion that the episode has to be followed by a contraction in output. Several studies that followed this used the same definition or modified the definition of extreme events used by Calvo, Izquierdo, and Mejia (2004). While Calvo, Izquierdo, and Talvi (2006) use the same definition, Calvo, Izquierdo, and Loo-Kung (2006) consider an additional criterion that these capital account reversals should be accompanied by an increase in some external aggregate measure of the cost of funds in order to capture systemic effects as such events are often accompanied by increase in systemic risk. Milesi-Ferretti and Razin (2000), Edwards (2007), and Adalet and Eichengreen (2007) analyze the causes and consequences of large and persistent current account reversals rather than focusing on short-run fluctuations. According to Milesi-Ferretti and Razin (2000), large current account reversals have to satisfy three criteria: the average current account deficit must fall by 2 (or 3)% of GDP between the first three and second three years; the maximum deficit in the second three years must be no larger than the minimum deficit in first three years; and the average deficit must fall by at least a third (as a percentage of GDP) between the first three and second three years. Reinhart and Reinhart (2009) analyze the other extreme of capital flow episode by looking at capital flow bonanzas, i.e., extreme inflows of net capital. They define bonanzas as periods when net inflows of a country exceed the top twentieth percentile of the country-specific distribution of net inflows. They also consider the current account deficit data as a proxy for net inflows data as it was available for a longer time period and for more countries as opposed to direct capital flows data. They 4

5 focus only on net inflows by imposing a non-negativity constraint on the net flows. 4 Cardarelli, Elekdag, and Kose (2010) study extreme inflows for a sample of 52 countries and define net private large capital inflows as percentage of GDP for a country as the ones which are higher than the country-specific rolling historical trend as well as higher than some region wide threshold where regions include Latin America, Emerging Asia, Emerging Europe and the Commonwealth of Independent States (CIS), an aggregate group of other emerging market countries and a group of advanced countries. Their region-wide threshold requires net private capital flows (in percentage of GDP) to be in the 75th percentile of the region wide distribution of the net capital flows (as percentage of GDP). Unlike earlier studies, they considered actual capital flows data and not current account deficit data as a proxy for the capital flows. They compute the flows data using various series from the IMF s International Financial Statistics, Balance of Payments and World Economic Outlook databases. Caballero (2012) defines net capital flows bonanzas as episodes when net inflows of a country grows more than during the normal business cycle and analyze their effect on the probability of a banking crisis. He decomposes the net flows data into trend and a cyclical component and define surges as periods when the deviation of the net inflows from the country-specific long-run trend is higher than one, two or half standard deviation of the cyclical component by some threshold level. A recent study by Powell and Tavella (2015) considers two definitions of capital flows surges using data on gross inflows for a panel of 44 emerging economies from 1980 to 2005 and analyze the effect of surges on the probability of banking crises and recessions. In their first definition, they define surges as deviations from a sample wide historical trend whereas in their second definition they define surges as episodes where gross inflows exceed one standard deviation above the country-specific trend. Again similar to Cardarelli, Elekdag, and Kose (2010) they compute the historical rolling trend using only past information by employing a Hodrick-Prescott filter. While Cardarelli, Elekdag, and Kose (2010) and Caballero (2012) focus on the macroeconomic consequences and policy implications of extreme events in net capital inflows, a recent study by Ghosh et al. (2014) analyze the plausible determinants of them for a sample of emerging market economies. Using annual data on actual capital flows they define surges as periods when the net capital flows belong to the top 30th percentile of the country-specific distribution as well as entire sample s 4 They do so to ensure the countries that have mostly outflows of capital on a net basis do not have any bonanzas. 5

6 distribution of net capital flows. 5 As opposed to the above studies that focus either on sharp increases in net inflows (surges or bonanzas) or sharp decreases in net inflows (sudden stops), Forbes and Warnock (2012) provide a more general study of different possible episodes in capital flows. They consider actual data on quarterly gross capital flows (inflows and outflows) and define four possible episodes: surges, stops, flights and retrenchments. The first two type of episodes are defined based on gross inflows whereas the last two are defined with regards to gross outflows. They consider gross as opposed to net flows in order to distinguish between the behavior of domestic and foreign investors. Given the residency-based definition of capital flows, surges and stops are driven by foreign investors and flights and retrenchments are driven by domestic investors. Their criterion for identifying these episodes is similar to that followed by Calvo, Izquierdo, and Mejia (2004). According to their definition, a surge episode begins if the value of the year over year changes in the four quarter sum of gross capital inflows increases one standard deviation above the historical mean 6 and ends when the value is below one standard deviation above the historical rolling mean. Also during the episode the annual change in four quarter sum of the gross inflow has to increase two standard deviations above the historical mean at least for a quarter anytime during the period. Analogously, they define a flight episode using gross outflows. The existing methodologies used in the literature identify such episodes by either focusing on the tail of the capital flows distribution by choosing a threshold for top percentile or by choosing values that deviate by some number of standard deviations away from the mean. The question that arises regarding such methodologies is how to choose the value of the threshold? Existing studies take this as given like Reinhart and Reinhart (2009) or Forbes and Warnock (2012). Ghosh et al. (2014), estimate a number of quantile regressions to show that large flows as indicated by higher quantiles of net flows distribution behave qualitatively differently on average from the rest of the distribution. However, they do not provide any justification of choosing the top 30th percentile as a threshold as opposed to, say, the top 20th or top 35th percentile. 5 The net capital flows considered by Ghosh et al. (2014) are in percentages of GDP. 6 The historical mean is calculated as a rolling mean over previous four years. 6

7 3 Methodology 3.1 Data I use quarterly data on net private capital flows from the first quarter of 1980 to the second quarter of 2014 for a sample of 36 emerging market economies for my analysis. 7 The net private capital flows series is computed using the net financial account excluding government liabilities from the IMF s Balance of Payment Statistics, which is similar to the definition followed by Ghosh et al. (2014). The details of the countries, data period, and the computation of the net flows series used can be found in Appendix B. According to the IMF s balance of payments (henceforth, BOP) accounting convention, a positive value of the flows indicates net inflows, i.e., capital flowing into the country on a net basis. Analogously a negative value of the flows indicates net capital outflow, i.e., capital flowing out of the country on a net basis. I control for the size of the economy by taking the net flows as a percentage of the country s nominal GDP as for larger economies a large flow may not be as much of a concern relative to a small economy as they may be better able to absorb such large flows. The nominal GDP data is obtained from the IMF s World Economic Outlook database. 8 Summary statistics of the data are provided in the Table 1. I have a total of 3703 quarterly observations for 36 emerging markets spanning up to 35 years from 1980 to I group the countries in the sample into four regions: Asia, Latin and Central America, Europe and CIS (Commonwealth of Independent States) and Other, where the Other group includes some African and Middle Eastern countries. 9 My sample consists of 10 Asian, 11 European and CIS and 12 Latin and Central American countries, and the Other group contains 3 countries. The overall sample mean of the net capital flows as a percentage of GDP is 0.6%. The East European and CIS countries have a higher mean of around 2.1% of GDP compared to other regions in the sample. Countries belonging to Asia, Latin and Central America and other African countries and Israel have means closer to each other of about 0.4%. Also there is quite a bit of variation in the minimum and maximum values of the flows across different regions. Figure 1 plots the average net 7 The choice of the countries and data period is restricted mostly by availability and quality of the data 8 I use the October 2014 version of World Economic Outlook. 9 In particular, Other includes Israel, Morocco and South Africa. 7

8 capital flows as a percentage of GDP for individual countries in the sample. Most of the countries in the sample have on an average net inflows of capital as evident from the positive values of the mean of net capital flows as percentage of GDP. There are five countries in the sample (Argentina, Bangladesh, Ecuador, Russia, and Venezuela) which experienced a net outflow of capital on an average over the entire sample period. Table 2 provides summary statistics of net capital flows in percentage of GDP for individual countries. Ecuador has the highest net outflow (-30.7%) and Latvia has the highest net inflow (9.5%) in percentage of GDP relative to other countries in the sample over the sample period of 35 years. 3.2 The Model The purpose of this paper is to identify periods of extreme flows for a sample of emerging market economies from the data. As mentioned above, a positive value of the flows indicates net inflows and a negative value indicates a net outflow of capital. Thus a country can have periods of extreme inflows or periods of extreme outflows. The period of extreme inflows would imply a very high positive value and a period of extreme outflows would imply a very high negative value of the capital flows. The question then is how to define such extreme values of the flows. I contribute to the literature by using a rigorous statistical model for identifying such extreme episodes solely from the data and not, say, looking at the consequences of the flows. As mentioned earlier, I use the Markov-switching model of Hamilton (1989) to characterize periods of extreme inflows and outflows. Given the episodic nature of capital flows to emerging market economies and with the objective of identifying exceptionally large flows, I assume that the net flows in a country follow a regime switching process with the regimes defining normal times, as well as times of extreme inflows and outflows, and times of high inflows and outflows. In particular, I allow the means of the absolute values of the net flows to evolve according to a three state Markov-switching process, thus allowing the dynamics of extreme flows to be qualitatively distinct from those of non-extreme flows (i.e., high and low flows). The switches between different states in this model are treated as inherent to the datagenerating process as opposed to the arbitrary threshold model. The obvious question is why use a three state model as opposed to a simpler two state model? My objective is to identify periods of net capital flows as percentage of GDP when they are exceptionally high and not just a period of high and low net flows. A two state model would be more likely to have difficulty distinguishing 8

9 capital flows episodes that are merely above average from those are truly extreme. I am interested in identifying periods of extreme flows both in and out of the country. In addition, I consider absolute values of the series in the Markov-switching model as opposed to the actual values of the flows to identify the periods of extreme flows. As I am considering net flows data, which implies values can be negative (indicating inflows) or positive (indicating outflows), application of a three state regime switching model in this case identifies only periods of net inflows, net outflows and flows that are close to zero on a net basis. An alternative would be to allow for six regimes in the Markov-switching model with actual net flows data. Here, the assumption would be that the net flows of the data switches between six regimes: very high, high and low values of net inflows and outflows respectively. However, increasing the number of regimes make the identification of the states less precise due to the limited amount of data I have for these emerging market economies. Using the absolute values helps me to identify the extreme states of the net flows of capital in percentage of GDP using a relatively parsimonious three state model. To understand the methodology, let y it denote the absolute values of the net capital flows (in percentage of GDP) for country i at time t. The model postulates that there exists an unobservable state variable (S t ) that can take values 1, 2 and 3 (as I consider a three state model). When S t is equal to 1, the absolute value of the flows for country i, y it is distributed normally with mean µ 1i and variance σi 2 (N(µ 1i, σi 2 )). Similarly when S t is equal to 2 and S t is equal to 3, the y it is distributed as N(µ 2i, σi 2 ) and N(µ 3i, σi 2 ) respectively. I assume that only the mean of the distribution switches across states. I define the regime with the highest value of the mean as the extreme state. Formally, I describe the model as: y it = µ Sit + ɛ it, (1) where ɛ it N(0, σi 2 ) The unobserved state variable S it follows a first order Markov process and is governed by the following transition probabilities: Pr[S t = i S t 1 = j] = p ij, i, j = 1, 2, 3 where i and j are states. I allow for switches only in the means of the distribution of the net capital flows keeping the variance unchanged under the different states because I am interested in identifying periods when the countries experience huge flows (inflows or outflows) compared to their average level. Hence, I do not allow the volatility to change. The parameters of the model that need to be estimated are µ 1, µ 2, µ 3, σ 2, 9

10 p 11, p 12, p 13, p 21, p 22, p 23, p 31, p 32, and p 33. Let the parameter vector be denoted by θ. It is estimated using maximum likelihood estimation using the Matlab Markov switching package developed by Perlin (2014). The details of estimation method for the parameter vector are explained in Hamilton (1989). I apply a smoothing algorithm for obtaining the state probabilities as proposed by Kim (1994). Also I estimate the above model for each country separately. Hence, the states of the capital flows for each countries are determined by country-specific data generating process. This gives me a parameter vector θ i for each country i. The three state Markov-switching model identifies regimes of extreme, high, and low flows for each country. The Markov switching model identifies regimes, and not observations. The model uses the parameter estimates to probabilistically classify observations in different regimes. If the probability of an observation being in the extreme regime is the highest among the three regimes, and the value of the net flow data is positive, I classify it as an extreme inflow. 10 Analogously if the probability of being in the extreme regime is the highest among the three regimes, and the net flow has a negative sign, it gets classified as an extreme outflow. In the literature, periods of extreme net inflows and net outflows are usually termed as surges and flights respectively. In this paper, I also employ that terminology and call the extreme inflows as surges and extreme outflows as flights. Similarly for the other two regimes (high and low flow regimes) I have outflows and inflows episodes separately. So I get six states for the net flows data using this methodology: surge, flight, high inflow, high outflow, low inflow, low outflow. 4 Results 4.1 Markov-Switching: Summary of Results Table 3 summarizes the different episodes identified by the three state Markovswitching model as described in Section 3. As mentioned earlier, the episodes are identified for each individual country separately. Table 3 aggregates across countries and reports summary statistics on the occurrences of different episodes, average value of flows, and the average duration of each episode. The model identifies a total of 289 quarters as surges and 110 quarters as flights out of the total 3703 observations. As evident from Table 3, there are relatively few flight episodes compared to surge episodes. The surges constitute around 8% of the total sample whereas flights are 10 I assume there is equal likelihood of an observation being in either of the three states. 10

11 around 3% of the total sample of observations. This is in line with earlier findings. Historically, the emerging market economies have experienced more episodes of wildly fluctuating inflows than outflows (Reinhart and Reinhart, 2009). Comparing the mean values of the flows across different episodes, it is evident that the extreme episodes do exhibit really high levels of net capital flows as a percentage of GDP. The surges on average have net capital flows of 3.4% of GDP and the flights involve net capital flows of 3.5% of GDP. The high inflows are about 1.5% of GDP and the low inflows are 0.75% of GDP on an average across all countries in the sample. The average flows in the surge state are about twice the value of average flows in the high inflow state and about four times the value in the low inflow state. The outflows in extreme periods are also double the size of high outflows and six times the value in the low outflow period. The mean flow in a high outflow state is -1.18% of GDP and in low outflow is -0.56% of GDP. The table also gives the expected duration of each regime measured in quarters which are obtained from the transition probabilities of the three state Markovswitching model. The expected duration (E(D)) of a regime, say j, is calculated as E(D) = 1 1 p jj, (2) where p jj is the probability of being in state j in period t conditional on being in state j in period t 1, j=1, 2, 3. Looking at the duration of the different episodes, it can be seen that the extreme episodes tend to be short-lived compared to the high and low flows periods. The surges on average tend to last for three quarters and the flights last for four quarters. The high flows periods are much more persistent lasting for about four years. The low net flows are a little shorter than the high flows period. The low inflows last eleven quarters on average and the low outflows last for thirteen quarters on average. 4.2 Surges and Flights In this section, I analyze the results for extreme flow episodes more closely by focusing on the surge and flight episodes for the different countries and different region groups. As mentioned in subsection 4.1, the three state Markov-switching model identifies 289 periods of surges and 110 periods of flights of net capital flows. I plot the time series of surges and flights in Figure 2. The top panel of the figure plots the surge episodes as a percentage of the total number of cross-sectional observations in each year on the right axis. On the left axis, the graph plots the average net capital flows as 11

12 a percentage of GDP for all the countries in the surge episode. The bottom panel on the other hand plots the flight episodes as a percentage of the total number of cross-sectional observations in each year on the right axis. Again on the left axis, the figure plots the average net capital flows in percentage of GDP for all the countries experiencing a flight episode. Although I consider a quarterly frequency of data to identify the surge and flight episodes for each country, in these graphs I aggregate the number of episodes to an annual level. From the top panel in Figure 2, we can see that the surge episodes as percentage of total cross-sectional observations have fluctuated considerably over the period of analysis. It was around 10% of the total cross-sectional observations in 1980 and then declined to zero in There were no surge episodes between 1984 and From 1989 onwards the surges of capital to emerging markets started to increase and they continued to increase until the mid-1990s after which again there was a sharp decline in the surge episodes. The number remained low in the early 2000s. From 2003 onwards, there was a steady increase in surge episodes in these economies peaking at 20% in 2007, right before the global financial crisis. This period corresponds to the house price boom in the United States. It was argued by Taylor (2007) that the Fed followed a loose monetary policy during the period from 2003 to Thus these emerging markets attracted foreign investors looking for higher return on their investments and an opportunity to re-balance their portfolio. The occurrences of surges again declined sharply in the wake of the global financial crisis of With the U.S. interest rate hit the zero lower bound after the financial crisis capital started surging towards the emerging markets as indicated by the surge occurrences as percentage of total observations from 2009 onwards. But the recent talk about exiting the unconventional policy regimes by the Federal Reserve Bank have triggered capital to flow out of these emerging markets. This can be seen in the graph as well. In 2013, the tapering talk by the Federal Reserve Bank, i.e., when the Federal Reserve was considering about tapering its asset purchases, created a taper tantrum and capital started flowing out of these economies (see (Eichengreen and Gupta, 2014)). The incidence of capital flight is much less than that of capital surges. But again, the occurrences of flight episodes have also fluctuated over the sample period. Comparing the top and bottom panel, it can be seen that the surge and flight incidences are mirror images of one another. The periods of surges correspond to periods of low outflows (no flight). When more countries are experiencing surges, correspondingly fewer are experiencing flights. Between 1984 and 1989 when there were no episodes 12

13 of surges, there were incidences of capital flight. There were around 4% of the total observations that experienced capital flights during this period. Between 1990 and 1993 there were no countries experiencing any extreme outflow of capital. But the percentage of observations belonging to surge period increased from 3% to 10%. If we look at the period of global financial crisis, there was an increase in the incidences of flight episodes at this time. Again with the reduction in advanced nations interest rates following the crisis, incidences of flight went down in these emerging market nations as they served as the recipients of the capital that was flowing out of the advanced nations. As mentioned in Section 3, I divide the entire sample of countries into four regions: Latin and Central America, Asia, Europe and CIS and Other. Figures 3 and 4 plot the surge and flight incidences respectively for different countries in the sample split by the region groups. Again I plot the average net capital flows as a percentage of GDP in surges and flights respectively on the left axis of the two figures for all region groups. It can be seen that the episodes identified by the Markov-switching (henceforth, MS) model match some well known historical events. For example, the MS model identifies periods of surges in Asian countries before the Asian Financial Crisis and periods of flights in the late 1990s and early 2000s. For Latin and Central American countries, the MS model identifies high incidence of flights between 2003 and 2006 which corresponds to the period after the South American economic crisis when Argentina, Brazil and Uruguay experienced some economic disturbances. Also the MS model is able to identify the capital outflows from Latin American countries after the Latin American Debt Crisis of the 1980 s as evident from the top left panel of Figures 4. Comparing across all four groups in Figure 3, it can be seen that the incidences of surges are not synchronized across all region groups. For example, Latin and Central American countries experienced surges more in the late 1990s whereas there were high incidence of surges for Asian countries in the early 1990s. The European and CIS countries experienced more surges in the mid 2000s. The Asian, European and Other countries experienced a steady increase in the incidences of surges in the 2000s until the onset of the Great Recession. But the Latin and Central American countries fluctuated a lot in this period. Looking at Figure 4, we see that the global financial crisis did not affect all the regions to the same extent. There was a sharp increase in the incidences of flight episodes in the European region relative to other groups. Asia also had an increase in the percentages of flight episodes relative to total observations in the region but it was much less compared to what the Asian economies 13

14 had experienced after the Asian financial crisis in In the Latin American countries, however, the flight incidences actually went down during the crisis. In order to understand further how well the MS approach identifies the extreme episodes in the different countries, I look at some of the individual countries in the sample. Figure 6 plots the net capital flows of eight countries: Argentina, Mexico, Thailand, Indonesia, Brazil, Russia, India, and South Africa. For Argentina, the MS model identifies periods of flights in the late 1980s, i.e., after the Latin American debt crisis, and a few periods of flights after the depression in the country. For Mexico, the MS approach identifies periods of surges before the Mexican Peso Crisis of Also there is some evidence of a surge in capital in 2009, when the U.S. interest rate hit the zero lower bound. The graph shows that there was some capital leaving the Mexican economy after the devaluation of the Mexican peso in December However, the MS model results show that this was not high enough to be part of a different data generating process and hence, these periods are not classified as extreme capital outflows. For Indonesia and Thailand, again the MS model captures episodes of flights around the period of Asian Financial Crisis of For Indonesia, the capital flight episode started in 1997:Q4 and continued till 2001:Q3. Next, I analyze the results for four major emerging market economies, Brazil, Russia, India and South Africa, which are also part of the BRICS group. 11 According to the MS model, Brazil experienced only one period of surge in the second quarter of Only India and South Africa experienced surges in capital flows in the post Global Financial Crisis period when the Fed lowered the interest rates in U.S. Also the results indicate that the tapering talk by the Fed in the summer of 2013 did not lead to a significant outflow of capital for any of these countries. Russia experienced two episodes of flights, one in the third quarter of 2000 and the other in the fourth quarter of It is experiencing outflows of capital in recent years, however, these outflows have not been large enough to be classified as distinct flight episodes by the MS model. 5 Comparison with Existing Methods In this section, I compare the results that I get from using the three state Markovswitching model with the methods used in the current literature. As mentioned ear- 11 BRICS is an acronym for Brazil, Russia, India, China and South Africa that refer to an association of these major emerging market economies. 14

15 lier, the literature on analysis of net capital flows tends to rely on arbitrary threshold approaches to identify periods of extreme flows. The study that comes closest to my analysis is that of Ghosh et al. (2014) as the data on net capital flows I use is very similar to theirs. 12 However, the sample of countries and the frequency of the data I consider is different. In their paper, they identify periods of net inflows that are abnormally high using data on annual net private capital inflows as a percentage of GDP for a sample for 56 emerging market economies from 1980 to According to their methodology a period (measured in years) is considered to be a surge episode if it belongs to the top 30th percentile of country s own distribution of net capital inflows as well as in the top 30th percentile of the entire sample distribution of net capital flows. Since the data and the sample of countries I consider do not exactly match with that used by Ghosh et al. (2014), in order to compare my results to theirs, I apply their methodology to my data to identify the surge periods. 13 Table 7 provides a comparison of surge episodes between the Markov-switching model (henceforth, MS) and the threshold approach used by Ghosh et al. (2014) by aggregating across all countries in the sample. It should be noted that their method identifies only surge episodes as they consider the top 30th percentile of net flows of capital where a positive value indicates a net inflow of capital. The model that I use, however, characterizes simultaneously both periods of extreme inflows and outflows on a net basis. But for comparing with the threshold approach used by Ghosh et al. (2014), I consider only the surge episodes identified by the Markov-switching model. Table 7 reports the surge episodes identified by the Markov-switching model as the horizontal variable and the surges identified using the 30th percentile threshold approach as the vertical variable. As discussed in section 4.1, the MS model identifies only 7.8% of the total observations in the sample as surges (a total of 289 surge episodes). 14 Table 7 shows that the threshold approach identifies 849 quarters as surge episodes (22.3% of the total sample) for the same sample of countries which is almost thrice the number identified by the MS approach. There are 263 surge episodes that are identified under both the the methods, which is around 91% of the 12 Forbes and Warnock (2012) also identify periods of extreme flows using actual flows data. However, their study analyzes gross flows as opposed to net flows. Also the criterion they use for identifying the extreme flows is based on a threshold in the standard deviation of the flows, i.e., it is more of a volatility threshold. 13 Reinhart and Reinhart (2009) also use a threshold approach to date episodes of extreme capital inflows, which they refer to as bonanzas. However, they used current account data as a proxy for net capital flows data rather than using actual net flows data. They considered a threshold of 20th percentile for all countries to identify the capital inflow bonanzas. 14 The episodes are measured in number of quarters 15

16 total episodes identified under the MS approach. However, 26 episodes out of the 289 episodes identified by the MS approach are not identified by the threshold approach. The correlation between the two methods is It should, however, be noted that the MS algorithm classifies surges for each country separately whereas the threshold approach followed by Ghosh et al. (2014) involves a country-specific threshold as well as a sample-wide threshold. Due to this there can be some differences in the identification of the surge episodes between the two methods. In their paper, Ghosh et al. (2014) also consider a country-specific threshold of the top 30th percentile as a robustness check. I also make a comparison of my results with the country-specific cutoff for net capital flows as percentage of GDP. The results of the comparison of the two methods are summarized in Table 8. By changing the cutoff to a country-specific distribution of net capital flows, the number of surges increases to This is not surprising as now around 30 percent of the total sample will be regarded as surges. Out of the 289 surges identified by the MS model, now 272 (94%) of them matches with the episodes identified by the country-specific threshold criterion. Although for the overall sample, I find a high match rate for the surge episodes between the two algorithms, however, there seems to be quite a bit of heterogeneity in the match rates across different countries. This can be seen from Figure 5 and Table 5. The threshold algorithm used by Ghosh et al. (2014) identifies on average more periods as surges compared to the MS algorithm. There are two countries in the sample, Hungary and Sri Lanka, for which the MS model identifies more periods of surges compared to the threshold approach. For most of the countries the MS algorithm identifies much fewer surge episodes than the threshold approach. For example Argentina, the MS model identifies only one quarter as the surge period whereas the threshold approach identifies 13 quarters as surges. For Albania and South Africa, however, the number of surges identified by the two methodologies are close. The threshold approach identifies 24 quarters as surges for Albania and 23 quarters as surges for South Africa. The MS model identifies 21 quarters and 20 quarters as surges for Albania and South Africa respectively. For South Africa all these 20 periods are identified by the threshold approach as well as MS model as evident from the third column of Table 5. However, for Albania 16 out of the 21 quarters are identified as surges under the threshold approach. For a better understanding of how the surge episodes vary across the two methods, consider the Figures 7, 8, 9 and 10 In these graphs I plot the net capital flows of each country as a percentage of GDP from 1980:Q1 to 2014:Q2 and highlight the periods of 16

17 surges and flights as identified by the three state MS model. 15 The gray shaded bars in the graphs indicate the periods of surges and the green shaded areas highlight the periods of flights as identified by the MS model. The red line represent the threshold criterion used by Ghosh et al. (2014). Recall that according to their algorithm a surge episode for a country not only needs to be in the top 30th percentile of its own distribution of net capital flows (in percentage of GDP) but also has to belong to the top 30th percentile of the entire sample s distribution of net capital flows (in percentage of GDP). This implies for the countries with a top 30th percentile value higher than the sample wide value of the top 30th percentile, the effective criterion for being qualified as surge episode will be its own country-specific 30th percentile. Similarly for countries with a lower value of the top 30th percentile of net capital flows relative to the sample wide value, the effective criterion for identifying surges is the top 30th percentile of the entire sample. So the red line in the graph represents the effective criterion (either the country-specific or the sample wide top 30th percentile value) to identify surge episodes used by Ghosh et al. (2014). I plot these graphs separately for all the 36 countries in the sample. Again I divide the sample into four groups according to the region: Latin and Central America, Asia, Europe and CIS, and the other countries are grouped into Other. By looking at the graphs in Figures 7, 8, 9 and 10, it is evident that the MS algorithm picks up periods as surges and flights in which the net capital flows as percentage of GDP is highly positive and highly negative respectively relative to the other periods in the sample which was the objective of this paper. There is quite a bit of variation in the duration of these episodes across countries. For a lot of countries, the surges and the flight tend to be short-lived, by lasting only for a quarter (for example Brazil, Colombia, and Korea). On the other hand, in countries like Sri Lanka, Hungary, Latvia and Ukraine the extreme episodes tend to be relatively more persistent (see Figures 8h, 9f, 9e, 9k). Ecuador, Russia, Korea, Poland, and Indonesia do not have any surges in net capital flows but rather only flight episodes. But based on the threshold criteria there are surges in these countries as depicted in Figures 9i, 8d, 9g, 8c. If we look at the graph for Argentina in Figure 7a, we find there is an exceptionally huge inflow in the third quarter of 1993 which gets tagged as a surge episode under the MS model but the threshold approach identifies lot of quarters between 1993:Q1 and 1999:Q4 15 As mentioned earlier, data for all countries are not available from 1980:Q1. So I consider the starting period for the different countries as the period when the data is available continuously. For most of the countries the sample period is until 2014:Q2. 17

18 as surges. We get a similar comparison for Brazil as well. For Brazil, the net capital flows in percentage of GDP went up to 8% in the first quarter of 1994 as evident from Figure 7b. Again, the MS model identifies only this quarter as an extreme inflow episode. However, under the threshold approach a lot of other observations get identified which have net flows values in percentage of GDP much lower than that of the first quarter of In some cases, the threshold criteria picks up episodes that are small fluctuations in net capital flows. For Morocco, on the other hand the two algorithms give exactly the same results (see Figure 10b). For Bangladesh, none of the periods qualify as surges based on the threshold criteria. The effective threshold criteria is the samplewide criteria for Bangladesh. However, the MS model identifies 6 periods as surges. This shows the importance of having country-wise criteria for identification of extreme episodes. One of the plausible reasons why the two models give such different results is due to the way the MS algorithm works. The MS model assumes the unobserved state variable follows a first order Markov process. This implies that the evolution of the state variable depends on the last period s state. Thus the probability that an observation is in an extreme state today depends on what the state was in the last period. However, the threshold methodology for identifying extreme episodes does not consider such dependence. It only uses rank information in the net capital flows distribution for classifying surges. As mentioned above, the threshold approach identifies around 23% of the total sample of observations as surges whereas the MS method identifies only 7.8% as surges. By changing the threshold level to 12.5% from 30%, then we get the same number of surges as the MS method, i.e., 289 quarters. However, only 164 out of these 289 observations are identified by the MS method which implies that by tightening the threshold criteria, although we are able to match the surge episodes under the two methods quantitatively, but qualitatively, they are still very different. The correlation between the two method is Conclusion This paper provides a formal methodology for identifying periods of extreme capital flows in emerging markets. Rather than relying on arbitrary thresholds, as previous studies have done, this paper employs a rigorous statistical model, namely, Hamilton (1989) s Markov switching model, to classify capital flow episodes entirely from the 18

19 data. Using quarterly data on net private flows for a sample of 36 emerging market economies over a period of 35 years from 1980 to 2014, the model identifies 7.8% of total observations as surges (extreme inflows) and 3% of total sample observations as flights (extreme outflows). The model is able to identify the states distinctly with the means of each state being statistically significant for each country in the sample. These extreme episodes tend to be short-lived compared to the other states of the net flows. However, there is a lot of variation in incidences of surges and duration of these extreme flows across countries. In comparison to the exogenous threshold criteria for identifying surges used in the literature, the Markov-switching model identifies a much lower incidence of surges. Even restricting the threshold criteria to match the number of surges that the Markovswitching model identifies, I find the surge episodes identified under the two methods differ considerably. Identification of the extreme episodes are crucial for conducting analyses on these extreme episodes in capital flows. Given the difference between the classification of episodes using a formal statistical model and using arbitrary threshold criteria is non-trivial, it may have important implications in the analyses of cause and effects of these episodes. This in turn may affect the policy implications based on the analysis of extreme events in capital flows. This paper does not analyze the determinants of these episodes or their consequences on the country s economies. The next step would be to take these episodes identified using the formal statistical methodology proposed in this paper, and analyze their determinants and consequences on countries s economies. 19